Using Machine Learning for Cosmetics and Cosmetic Ingredients
Transcript of Using Machine Learning for Cosmetics and Cosmetic Ingredients
![Page 1: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/1.jpg)
Using Machine Learning for Cosmetics and Cosmetic Ingredients
Timothy E. H. Allen
MRC Toxicology Unit, University of CambridgeDepartment of Chemistry, University of Cambridge
![Page 2: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/2.jpg)
Conflict of Interest Statement
I have no real or perceived conflicts of interest with the research described in this presentation.
![Page 3: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/3.jpg)
Outline/Objectives
To introduce how Unilever is applying the principles of next generation risk
assessment (NGRA)
To discuss where predictive computational toxicology and machine learning fit in
To show outline some of the computational approaches that have been
developed and compare them
To show why machine learning efforts are good for some tasks
To identify where these research efforts are now going, and how that impacts
their use
![Page 4: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/4.jpg)
The World of Consumer Products
3Divisions
PERSONALCAREFOODS &
REFRESHMENT
HOMECAREWE OPERATE IN
![Page 5: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/5.jpg)
Use of Safety Information for Consumer Products Context…
– Classification and labelling Favours a cautious approach Hazard based rules Occupational focus
– Screening/product development Many potential lead chemicals Often only hazard prediction methods are used Performance of models is less critical Exposure may be considered by the use of threshold-based approaches e.g.,
TTC, DST, EBW, EcoTTC– Risk Assessment including actual exposure
Requires a high degree of accuracy Route and amount of exposure dictate the need for toxicology data High level of scrutiny (internal and external)
Hazard Exposure
C2Cl3
ClClC
C2Cl3
ClClC
C2Cl3
ClClC
Cl
ClClCl
Cl
Cl
ClCl
ClClCl
Cl
ClClCl
Cl
Cl
Cl Cl
ClClCl
Cl
ClOHOH
OH
Lead Candidates
Marketed Chemical
Efficacy
Safety
Supply
![Page 6: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/6.jpg)
Ab Initio NGRA Framework
Baltazar, M.T., et al. (2020) Toxicol. Sci., Accepted Manuscript.
![Page 7: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/7.jpg)
Exposure estimate
Hazard estimate
Expo
sure
/ D
ose
Hazard ExposureMargin of Safety
![Page 8: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/8.jpg)
Exposure estimate
Hazard estimate
Expo
sure
/ D
ose
Hazard Exposure
![Page 9: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/9.jpg)
ICCR Principles of Risk Assessment without Animal Tests
The overall goal is human safety risk assessment The assessment is exposure led The assessment is hypothesis driven The assessment is designed to prevent harm
(i.e., distinguish between adaptation and adversity) Using a tiered and iterative approach Following an appropriate appraisal of existing information Using robust and relevant methods and strategies The logic of the approach should be transparently and explicitly
documented Sources of uncertainty should be characterised and documented
![Page 10: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/10.jpg)
Adverse Outcome Pathway
Ankley, G.T., et al. (2010) Environ. Toxicol. Chem., 29; 730.
![Page 11: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/11.jpg)
Adverse Outcome Pathway
Ankley, G.T., et al. (2010) Environ. Toxicol. Chem., 29; 730.
![Page 12: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/12.jpg)
Structural Alerts
Novel Compound
10No Hit Hit
Allen, T.E.H. et al. (2018) Toxicol. Sci., 165; 213.Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 13: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/13.jpg)
Structural Alerts
Allen, T.E.H. et al. (2018) Toxicol. Sci., 165; 213.
![Page 14: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/14.jpg)
Expanded Target ListAcetylcholinesterase Serotonin 3a (5-HT3a) receptor Vascular endothelial growth factor receptor 1 Serine/threonine-protein kinase PAK 4
Adenosine A2a receptor Serotonin transporter Vascular endothelial growth factor receptor 3 Phosphodiesterase 4A
Alpha-2a adrenergic receptor Tyrosine-protein kinase LCK Tyrosine-protein kinase FYN Phosphodiesterase 5A
Androgen receptor Vasopressin V1a receptor Glycogen synthase kinase-3 beta PIP2 3-kinase catalytic subunit α
Beta-1 adrenergic receptor Type-1 angiotensin II receptor Histone deacetylase 3 Peroxisome proliferator-activated receptor γ
Beta-2 adrenergic receptor RAC-alpha serine/threonine-protein kinase Insulin-like growth factor 1 receptor Protein Tyr phosphatase non-receptor type 1
Delta opioid receptor Beta-secretase 1 Insulin receptor Protein Tyr phosphatase non-receptor type 11
Dopamine D1 receptor Cholinesterase Vascular endothelial growth factor receptor 2 Protein Tyr phosphatase non-receptor type 2
Dopamine D2 receptor Caspase-1 Leukotriene B4 receptor 1 RAF proto-oncogene Ser/Thr-protein kinase
Dopamine transporter Caspase-3 Tyrosine-protein kinase Lyn Retinoic acid receptor alpha
Endothelin receptor ET-A Caspase-8 Mitogen-activated protein kinase 1 Retinoic acid receptor beta
Glucocorticoid receptor Muscarinic acetylcholine receptor M5 Mitogen-activated protein kinase 9 Rho-ass. coiled-coil-containing protein kinase I
hERG Inhibitor of nuclear factor κ-B kinase subunit α MAP kinase-activated protein kinase 2 Ribosomal protein S6 kinase alpha-5
Histamine H1 receptor Macrophage colony-stimulating fac. 1 receptor Hepatocyte growth factor receptor NAD-dependent protein deacetylase sirtuin-2
Mu opioid receptor Casein kinase I isoform delta Matrix metalloproteinase-13 NAD-dependent protein deacetylase sirtuin-3
Muscarinic acetylcholine receptor M1 Endothelin B receptor Matrix metalloproteinase-2 Proto-oncogene tyrosine-protein kinase Src
Muscarinic acetylcholine receptor M2 Neutrophil elastase Matrix metalloproteinase-3 Substance-K receptor
Muscarinic acetylcholine receptor M3 Ephrin type-A receptor 2 Matrix metalloproteinase-9 Thromboxane A2 receptor
Norepinephrine transporter Fibroblast growth factor receptor 1 Serine/threonine-protein kinase NEK2 Tyrosine-protein kinase receptor TEK
Serotonin 2a (5-HT2a) receptor Peptidyl-prolyl cis-trans isomerase P2Y purinoceptor 1
![Page 15: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/15.jpg)
Neural Networks
![Page 16: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/16.jpg)
Average Model Performance
Training Data Validation Data Test Data
SE SP ACC MCC ROC-AUC SE SP ACC MCC ROC-AUC SE SP ACC MCC ROC-AUC
AVERAGE 92.1 96.5 95.8 0.901 0.99 86.9 93.2 92.5 0.822 0.96 86.2 92.9 92.2 0.814 0.96
SD 8.8 4.2 3.1 0.069 0.02 11.7 5.9 4.1 0.091 0.04 12.1 6.5 4.2 0.093 0.04
![Page 17: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/17.jpg)
Model Performance vs. Total Compounds
![Page 18: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/18.jpg)
Positive Probability Curve
![Page 19: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/19.jpg)
Comparative Model PerformanceTarget Gene SE ΔSA ΔRF SP ΔSA ΔRF ACC ΔSA ΔRF MCC ΔSA ΔRF
AChE 88.4 6.6 -3.2 85.3 -4.6 7.5 87.0 1.6 1.6 0.737 0.024 0.030ADORA2A 97.6 2.8 -0.5 93.2 2.0 4.2 96.1 2.5 1.1 0.912 0.054 0.024ADRA2A 91.3 11.2 2.0 93.9 -1.2 -1.2 92.7 4.3 0.2 0.853 0.084 0.004
AR 66.5 -0.8 1.1 99.1 1.4 1.2 90.5 0.8 1.2 0.749 0.026 0.036ADRB1 92.7 7.0 0.3 89.5 -2.5 1.5 91.2 2.5 0.9 0.823 0.046 0.017ADRB2 72.9 -2.3 -3.3 89.9 2.2 1.4 81.6 0.0 -0.9 0.639 0.004 -0.014OPRD1 97.1 1.1 -1.2 81.0 -0.3 4.5 92.4 0.7 0.4 0.813 0.018 0.011DRD1 77.4 0.9 -3.7 96.6 1.5 4.3 89.2 1.4 1.2 0.773 0.030 0.028DRD2 98.3 1.9 -1.0 84.8 5.7 7.8 96.1 2.5 0.5 0.855 0.091 0.021
SLC6A3 89.9 1.3 -3.1 94.5 1.9 4.9 91.8 1.5 0.2 0.837 0.032 0.010EDNRA 93.8 -0.6 -3.4 95.6 1.7 4.8 94.6 0.4 0.4 0.893 0.010 0.009NR3C1 74.5 2.3 0.7 96.9 0.1 0.6 90.1 0.7 0.6 0.760 0.018 0.015KCNH2 84.4 15.7 -8.6 70.9 -11.5 17.4 79.1 5.0 1.6 0.558 0.059 0.036HRH1 95.2 8.0 -0.6 88.4 -5.3 0.7 92.0 1.7 0.0 0.840 0.032 -0.001
OPRM1 94.8 1.2 -1.1 94.5 1.7 3.2 94.7 1.4 0.6 0.889 0.030 0.014CHRM1 96.6 6.4 0.9 83.3 -2.9 0.7 91.7 2.9 0.8 0.821 0.062 0.019CHRM2 93.9 2.9 0.0 94.5 1.1 3.2 94.2 1.9 1.8 0.883 0.039 0.035CHRM3 91.9 3.4 -3.2 93.8 1.8 5.8 92.8 2.8 0.8 0.854 0.053 0.017
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 20: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/20.jpg)
Comparative Model PerformanceTarget Gene SE ΔSA ΔRF SP ΔSA ΔRF ACC ΔSA ΔRF MCC ΔSA ΔRF
SLC6A2 94.9 4.2 -0.2 92.5 -1.0 1.0 93.9 2.0 0.3 0.875 0.039 0.008HTR2A 99.1 2.4 -0.4 88.2 -0.4 6.1 96.7 1.8 1.0 0.901 0.051 0.030HTR3A 89.4 5.7 -1.0 98.2 -0.4 0.4 95.8 1.3 0.0 0.893 0.034 0.000SLC6A4 98.4 2.7 -0.3 89.0 2.1 6.3 96.3 2.5 1.2 0.892 0.070 0.037
LCK 95.5 3.7 0.2 79.8 -2.6 -0.9 92.3 2.5 0.0 0.763 0.054 -0.002AVPR1A 93.9 1.9 0.6 99.3 1.5 4.8 97.2 1.6 3.2 0.941 0.034 0.068AGTR1 87.3 0.0 3.4 99.3 1.0 1.9 94.5 0.6 2.6 0.888 0.014 0.054AKT1 95.4 1.0 -1.4 91.3 4.4 6.6 94.1 2.1 1.1 0.864 0.049 0.028
BACE1 92.0 -1.1 -5.7 93.5 5.7 14.1 92.5 0.9 0.1 0.827 0.028 0.016BCHE 85.6 7.1 -0.9 93.6 0.6 4.7 90.4 3.3 2.5 0.799 0.067 0.048CASP1 69.1 5.3 -1.9 94.7 -0.1 -0.1 86.5 1.5 -0.8 0.680 0.039 -0.018CASP3 84.8 3.4 3.7 94.9 -1.1 -1.1 91.0 0.6 0.7 0.809 0.013 0.015CASP8 86.8 -4.9 -2.5 95.4 -2.1 -2.8 92.1 -4.1 -4.1 0.832 -0.060 -0.059CHRM5 87.4 3.6 -4.2 95.3 1.5 4.0 92.3 2.3 0.9 0.835 0.048 0.015CHUK 88.8 5.7 -3.3 97.4 0.4 0.0 95.2 1.7 -0.9 0.871 0.047 -0.024CSF1R 94.3 5.9 -2.4 97.0 0.8 3.1 95.5 3.7 0.0 0.910 0.070 0.001
CSNK1D 91.4 12.4 3.8 94.8 -1.6 2.0 93.4 4.4 2.8 0.864 0.085 0.056EDNRB 96.1 2.0 -0.5 94.4 -0.4 -0.7 95.1 0.6 -0.6 0.899 0.013 -0.012
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 21: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/21.jpg)
Comparative Model PerformanceTarget Gene SE ΔSA ΔRF SP ΔSA ΔRF ACC ΔSA ΔRF MCC ΔSA ΔRF
ELANE 90.9 -0.4 -5.2 93.8 0.2 5.8 92.1 -0.1 -0.8 0.839 -0.001 -0.012EPHA2 87.2 0.0 -0.7 99.6 1.1 1.8 95.4 0.7 1.0 0.899 0.018 0.023FGFR1 96.8 5.9 0.8 92.0 -2.8 1.4 95.1 2.8 1.0 0.892 0.054 0.021
FKBP1A 88.1 -3.0 -6.9 97.4 -0.4 1.5 94.8 -1.1 -0.9 0.869 -0.028 -0.025FLT1 91.6 1.6 -3.0 99.0 0.9 2.5 96.2 1.2 0.4 0.919 0.024 0.008FLT4 91.9 10.7 0.0 96.5 -0.4 1.2 94.5 4.3 0.6 0.888 0.086 0.014FYN 77.8 5.1 -3.0 98.2 0.4 1.5 92.7 1.6 0.3 0.809 0.044 0.006
GSK3B 96.8 12.2 -1.1 78.3 -5.2 2.5 90.4 6.2 0.1 0.785 0.121 0.001HDAC3 94.1 2.3 0.3 94.8 1.5 2.7 94.5 2.0 1.6 0.890 0.039 0.032IGF1R 94.6 -1.0 -2.1 95.5 1.3 5.1 94.9 -0.2 0.3 0.889 -0.003 0.011INSR 91.7 1.7 -3.1 98.9 1.2 1.9 95.5 1.4 -0.5 0.912 0.028 -0.007KDR 97.4 2.8 -0.3 73.1 -9.7 9.2 93.5 0.8 1.2 0.748 0.005 0.058
LTB4R 91.4 4.3 2.2 98.8 0.0 1.2 96.8 1.2 1.5 0.918 0.030 0.038LYN 89.1 10.9 -1.8 98.0 2.7 2.7 95.4 5.2 1.4 0.888 0.127 0.030
MAPK1 43.5 -4.4 2.3 99.2 4.0 -0.2 79.3 1.0 0.6 0.557 0.043 0.013MAPK9 95.5 3.5 -2.9 97.7 2.2 7.9 96.5 2.9 2.0 0.931 0.058 0.040
MAPKAPK2 86.9 5.1 -3.3 94.1 1.4 2.5 91.0 3.0 0.0 0.816 0.062 -0.001MET 97.7 4.3 -1.0 91.9 3.2 8.4 95.9 4.0 1.9 0.904 0.091 0.045
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 22: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/22.jpg)
Comparative Model PerformanceTarget Gene SE ΔSA ΔRF SP ΔSA ΔRF ACC ΔSA ΔRF MCC ΔSA ΔRF
MMP13 94.9 1.5 -2.1 94.7 3.3 10.6 94.8 2.1 2.3 0.887 0.045 0.054MMP2 95.5 1.7 0.0 89.2 -3.6 6.0 93.2 -0.2 2.1 0.852 -0.006 0.048MMP3 94.7 1.8 -1.7 92.4 1.2 8.8 93.8 1.5 2.1 0.868 0.033 0.047MMP9 81.6 -0.2 -5.0 88.8 -1.5 6.9 84.6 -0.8 0.0 0.696 -0.016 0.011NEK2 77.0 9.4 -2.7 98.5 0.7 1.0 94.0 2.6 0.3 0.813 0.085 0.007P2RY1 92.7 -2.4 -2.4 100.0 1.5 3.4 97.7 0.3 1.5 0.947 0.007 0.035PAK4 89.4 7.0 -4.7 99.3 0.8 1.1 96.9 2.2 -0.3 0.914 0.064 -0.009
PDE4A 90.8 2.9 -1.1 94.9 -0.4 0.5 93.1 1.0 -0.3 0.859 0.020 -0.005PDE5A 90.1 3.4 -4.8 96.5 2.8 7.3 93.0 3.1 0.6 0.862 0.062 0.016PIK3CA 98.9 1.5 -0.4 93.3 -1.5 4.6 97.2 0.6 1.2 0.934 0.014 0.027PPARG 69.5 -0.8 -2.9 96.1 3.4 2.0 86.0 1.8 0.2 0.702 0.041 0.006PTPN1 76.6 9.5 -3.9 89.7 -2.6 3.1 84.7 2.0 0.4 0.673 0.046 0.005
PTPN11 64.8 26.2 14.8 91.9 -3.4 -3.4 85.7 3.4 0.8 0.584 0.153 0.052PTPN2 67.9 2.5 2.5 96.0 -1.7 2.3 90.1 -0.7 2.4 0.687 -0.022 0.069RAF1 99.7 4.2 0.0 95.2 -1.1 1.5 97.7 1.8 0.7 0.954 0.038 0.013RARA 63.1 1.9 1.9 99.4 1.8 0.0 95.2 1.9 0.3 0.742 0.092 0.013RARB 85.9 11.3 5.6 99.9 1.1 0.0 98.8 1.9 0.5 0.913 0.137 0.033
ROCK1 94.3 5.4 -2.5 92.0 -3.6 1.1 93.2 1.2 -0.9 0.864 0.021 -0.018
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 23: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/23.jpg)
Comparative Model PerformanceTarget Gene SE ΔSA ΔRF SP ΔSA ΔRF ACC ΔSA ΔRF MCC ΔSA ΔRF
RPS6KA5 73.7 12.3 0.0 100.0 1.5 2.7 95.3 3.4 2.2 0.835 0.135 0.080SIRT2 70.8 2.3 5.6 95.2 -0.6 -1.6 89.8 0.0 0.0 0.692 0.003 0.006SIRT3 76.7 -2.4 -4.7 98.5 0.8 0.8 95.4 0.3 0.0 0.802 0.010 -0.005SRC 94.7 4.6 -2.7 88.7 0.7 8.1 92.4 3.1 1.5 0.839 0.064 0.029
TACR2 87.6 -1.6 -4.4 100.0 2.4 2.4 96.0 1.1 0.2 0.910 0.029 0.008TBXA2R 88.8 -0.4 -2.1 94.9 -0.4 1.4 93.0 -0.4 0.3 0.838 -0.010 0.003
TEK 89.9 2.5 -4.1 97.8 2.6 1.9 94.5 2.6 -0.6 0.887 0.053 -0.013
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 24: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/24.jpg)
Neural Network Activation Similarity
![Page 25: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/25.jpg)
Amiodarone (hERG)
![Page 26: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/26.jpg)
Amiodarone (hERG)
![Page 27: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/27.jpg)
Amiodarone (hERG)
![Page 28: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/28.jpg)
Quantitative Predictions
Dose-Response Relationships and Risk Assessment Procedures require
Quantitative Information
Adjustment of AR dataset to contain only quantitative activity vales
(p(Activity), 4880 values)
Change of loss function to MSE
Single output node with linear activation function
Models evaluated using MSE and RMSE
![Page 29: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/29.jpg)
Quantitative Predictions
![Page 30: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/30.jpg)
Quantitative Predictions
![Page 31: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/31.jpg)
Summary
Neural Networks are a class of Machine Learning Algorithms that can provide
both binary and quantitative predictions
Structural Alerts, Random Forests and Neural Networks have been used to try
and predict binary activity at Human MIEs
A combination of these models and understanding of their workings is key to
highest performance and model use in toxicology decision making
Quantitative predictions help push this methodology closer to use in risk
assessment, rather than just hazard identification
![Page 32: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/32.jpg)
References
Adverse outcome pathways: A conceptual framework to support ecotoxicology research and risk assessment, Ankley, G.T., et al. (2010) Environ. Toxicol. Chem., 29; 730.
A next generation risk assessment case study for coumarin in cosmetic products, Baltazar, M.T. et al. (2020) Toxicol. Sci., Accepted Manuscript.
Using 2D Structural Alerts to Define Chemical Categories for Molecular Initiating Events, Allen, T.E.H. et al. (2018) Toxicol. Sci., 165; 213.
Structural Alerts and Random Forest Models in a Consensus Approach for Receptor Binding Molecular Initiating Events. Wedlake, A.J. et al. (2019) Chem. Res. Toxicol., 33; 388.
![Page 33: Using Machine Learning for Cosmetics and Cosmetic Ingredients](https://reader030.fdocuments.net/reader030/viewer/2022012723/61b538992078661f815d7d39/html5/thumbnails/33.jpg)
Acknowledgements
• Professor Jonathan Goodman
• Professor Anne Willis
• Unilever
• Dr. Paul Russell, Dr. Steve Gutsell & colleagues at SEAC, Unilever
• Dr. Andrew Wedlake
• Maria Folia and Dr. Sam Piechota
• The Centre for Molecular Informatics
• The MRC Toxicology Unit
• St. John’s College